Make Your Code Faster with Data Parallelism

Daniel Lemire, professor
Université du Québec (TÉLUQ)
Montréal

blog: https://lemire.me
X: @lemire
GitHub: https://github.com/lemire/

Where I am coming from

• Author of libraries you are probably running right now: simdjson, simdutf, fast_float, Roaring Bitmaps.
• Shipped inside .NET, Rust, Go, Node.js, Bun, Deno, Google Chrome, Safari, ClickHouse.
• Node.js core contributor (C++ review, performance, security).
• Editor, Software: Practice and Experience (Wiley, founded 1971).

Part 1

Parallelism for the win

Some numbers

• Time is discrete: the clock cycle
• Processors: ~4 GHz
• One cycle is 0.25 nanoseconds
• Light travels 7.5 cm per cycle
• One byte per cycle is only 4 GB/s

Clock frequency: 25 years

• Pentium 4 (2000): 1.3 to 2 GHz
• AMD Ryzen 7 9800X3D (2024): 4.7 to 5.2 GHz
• About 2.5× in 25 years

Clock frequency: AMD Ryzen 7, 2022–2024

• Up 15% is two years

Geekbench 6: AMD Ryzen 7, 2022–2024

Single-core: +47%. Multi-core: +58%. Same 8 cores.

Transistors: AMD Ryzen 7, 2022–2024

• 8 cores
• Up 50% is two years

+5.5 B transistors, 2022–2024

• Same 8 cores
• 76% went into the core die
• 24% into the I/O die
• Cache die unchanged

What changed in the core: SIMD (same 8 cores)

Same software, three years apart

• simdjson, PartialTweets benchmark, one core
• Same code: 2.4× from 2022 to 2024
• The library is designed for data parallelism
• Same $

Disk at gigabytes per second

You are CPU bound.

More than you know.

Zen 5 AWS (EPYC 9R45, c8a)

• STREAM single thread bandwidth: 46 GB/s
• strstr, 32-byte needle: 9.5 GB/s

Data-level parallelism

The rest of this talk

SIMD (Single Instruction, Multiple Data)

• Process 16, 32 or 64 bytes with one instruction
• Supported on every modern CPU — your phone included
• Not a niche feature

The instruction sets

• x64: SSE2 (baseline, 128-bit), AVX2 (256-bit), AVX-512 (512-bit)
• ARM: NEON (128-bit, baseline on ARMv8), SVE/SVE2 (scalable)
• RISC-V: RVV (scalable)
• POWER: AltiVec/VSX · LoongArch: LSX, LASX
• WebAssembly: 128-bit SIMD, in every browser

Portability: compile several kernels, dispatch at runtime on CPU features. C++26 adds data-parallel types (std::simd).

SWAR: SIMD within a register

• Use plain 64-bit integer instructions
• A 64-bit register holds 8 bytes
• Fully portable C, C++, Rust, Go, Java...
• No intrinsics, no dispatch, no #ifdef
• Requires some cleverness

A great place to start.

Check whether we have 8 digits

add     rax, rdi
add     rdi, rdx
or      rax, rdi
test    rax,

Case study: ASCII processing

ASCII to lower case

For each character c
    If c - 'A' <= 'Z' - 'A' then
        c = c + 'a' - 'A'
    EndIf
EndFor

One byte per iteration. One unpredictable branch per byte.

64 characters in 3 instructions

Let us check with llvm-mca

• llvm-mca is a static machine-code analyzer shipped with LLVM
• Feed it assembly, name a microarchitecture, get a cycle estimate
• No hardware required, no noise, no warm-up

llvm-mca -mcpu=icelake-server -iterations=100 kernel.s

The scalar loop: 1 byte per iteration

movzbl  (%rdi,%rax), %ecx
leal    -65(%rcx), %edx
cmpb    $26, %dl
jae     .LBB1_2
addb    $32, %cl
movb    %cl, (%rdi,%rax)
addq    $1, %rax
cmpq    %rax, %rsi
jne     .LBB1_1

The AVX-512 loop: 64 bytes per iteration

vmovdqu64  (%rdi,%rax), %zmm2
vpsubb     %zmm0, %zmm2, %zmm3
vpcmpub    $2, %zmm1, %zmm3, %k1
vpaddb     %zmm4, %zmm2, %zmm2 {%k1}
vmovdqu64  %zmm2, (%rdi,%rax)
addq       $64, %rax
cmpq       %rax, %rsi
jne        .LBB0_1

llvm-mca says...

About 60× fewer cycles per byte.

What llvm-mca will not tell you

• It assumes perfect branch prediction
• It assumes every load hits L1
• It often ignores memory alignment issues
• It models one microarchitecture at a time.

It is a model. Not reality.

Reality: Intel Xeon Gold 6548N (Emerald Rapids)

Alignment: 64-byte loads and stores, 64 offsets

Only 1 offset in 64 matches the model. The other 63 straddle two cache lines: +30%.

Masked loads and stores

• AVX-512 loads and stores take a 64-bit mask: one bit per byte
• Masked-out bytes are not read, not written, and cannot fault
• Round p down to the cache line, mask off the bytes before p: the rest of the loop is aligned
• Same trick handles the tail: no scalar loop, no branch

Case study: JSON

Joint work with many people such as Geoff Langdale (Intel), John Keiser (Microsoft), Francisco Geiman Thiesen (Microsoft), etc.

JSON

• Portable, simple, human-readable; used by ~97% of API requests
• Strings (escaped), numbers, objects, arrays
• Reading and writing JSON is often slow: 100 MB/s to 300 MB/s
• Slower than a fast disk or a fast network

$ go run parse_twitter.go
Parsed 0.63 GB in 6.961 seconds (90.72 MB/s)

You are probably using simdjson

• Node.js, Electron
• ClickHouse
• WatermelonDB, Apache Doris, Meta Velox, Milvus, QuestDB, StarRocks

Every major JavaScript engine parses JSON with SIMD

• JSON.parse in Node.js, Bun and Deno is data-parallel (simdjson inspired).
• Billions of calls per second worldwide.
• Nobody had to change a single line of JavaScript.

simdjson: two-stage design

Stage 1 (data-parallel): scan the whole document with SIMD

• find every structural character and the start of every string
• validate UTF-8
• produce an index

Stage 2 (mostly serial): walk the index and build values

• the hard, branchy work now runs on ~5% of the bytes

Deserialization (Intel Xeon Gold 6548N)

Serialization (Intel Xeon Gold 6548N)

Classifying characters

We need to sort every byte into a class:

• comma (0x2c) ,
• colon (0x3a) :
• brackets (0x5b, 0x5d, 0x7b, 0x7d): [, ], {, }
• white-space (0x09, 0x0a, 0x0d, 0x20)
• everything else

A switch statement per byte? No.

Classification, 64 bytes at a time

Serialization is also a data-parallel problem

• JSON requires escaping ", \, and control characters.
• Almost no string actually needs escaping.
• So: check 16 to 64 bytes at once, and take the fast path.

Case study: IP addresses

Joint work with Yagiz Nizipli (SpaceX)

192.168.0.1 → 32 bits

• Every server logs the client address. Every firewall evaluates a rule. Every URL parser meets an IP literal.
• The usual tool: inet_pton from the C library, ~340 instructions per address.
• An IPv4 address is at most 15 bytes: it fits in one 16-byte register.
• An IPv6 address is at most 45 bytes: it fits in one 64-byte register.

Load the whole address once. Never walk it.

Part 5

Why your compiler will not do this for you

After decades of autovectorization research

• C# (.NET) has intrinsics
• C++ added std::simd
• Java Vector

Questions?

Daniel Lemire — lemire.me

X: @lemire · GitHub: github.com/lemire

Case study: Unicode

simdutf

• Inside Safari, Chrome, Node.js, Bun
• Unicode transcoding and validation at gigabytes per second
• Base64 too
• x64, ARM, POWER, RISC-V, LoongArch

Unicode (UTF-16)

• Code points from U+0000 to U+FFFF: a single 16-bit value.
• Beyond: a surrogate pair, U+D800–U+DBFF followed by U+DC00–U+DFFF.
• A lone surrogate makes the string ill-formed.

Every JavaScript string, every Java string, every Windows filename.

Validate

PROCEDURE validate_utf16(code_units)
    i ← 0
    WHILE i < |code_units|
        unit ← code_units[i]
        IF unit ≤ 0xD7FF OR unit ≥ 0xE000 THEN
            INCREMENT i
            CONTINUE
        IF unit ≥ 0xD800 AND unit ≤ 0xDBFF THEN
            IF i + 1 ≥ |code_units| THEN
                RETURN false
            next_unit ← code_units[i + 1]
            IF next_unit < 0xDC00 OR next_unit > 0xDFFF THEN
                RETURN false
            i ← i + 2  // Valid surrogate pair
            CONTINUE
        RETURN false
    RETURN true

UTF-16 correction, Apple M4

13× fewer instructions per byte.

In the browser (Apple M4)

• Chromium: 16 GB/s (uses our new function)
• Firefox: 3.4 GB/s
• Safari: 1.2 GB/s

Test it yourself: https://lemire.github.io/browserwellformed/

Data parallelism applies more often than you think

• JSON parsing — every JavaScript engine
• Unicode validation — every browser
• base64 — the JavaScript standard library
• Number parsing — every compiler toolchain
• Bitmap indexes, hashing, compression, memchr, CSV, regex prefiltering

"Inherently serial" is usually a statement about the algorithm you happen to know.

Interested? Check these projects

• simdjson — the fastest JSON parser in the world https://simdjson.org
  • Node.js, Bun, Deno, Electron
  • ClickHouse, WatermelonDB, Apache Doris, Meta Velox, Milvus, QuestDB, StarRocks
• simdutf — Unicode (UTF-8/16/32) and base64 https://github.com/simdutf/simdutf
  • Node.js, Bun, WebKit (Safari), Chromium (Chrome, Edge)
• fast_float — number parsing https://github.com/fastfloat/fast_float
• Roaring Bitmaps — https://roaringbitmap.org

Credit

• simdjson reflection work with Francisco Geiman Thiesen (Microsoft)
• simdutf UTF-16 correction is joint work with Robert Clausecker
• simdjson and simdutf are community efforts (Geoff Langdale, John Keiser, Paul Dreik, Yagiz Nizipli and others)

Extra slides

Part 7

Measure properly, or do not bother

Measurements

• We often assume that measurements (timings) are normally distributed.
• If they were, the 'error' would fall off as .
• It is often an incorrect assumption.

What if we dealt with log-normal distributions?

Real-world measurements

• You cannot assume normality
• Measurements are not independent
• Reality: the absolute minimum is often the reliable metric
• Margin: the difference between the mean and the minimum

Use performance counters

Timings tell you that something is slow. Counters tell you why.

• instructions retired — did I actually remove work?
• cycles — the ground truth
• branch misses — is the predictor carrying me?
• cache misses — am I memory bound after all?

perf stat, Instruments, or a library such as performancecounters.

Hot-spot engineering fails

• A profiler shows you where the cycles are now.
• It does not show you the 30% of instructions spread evenly over every function.
• Reducing the total instruction count beats optimizing the top-of-profile function.

Part 6

Can an LLM write your SIMD code?

Partly

Frontier models in 2026 are genuinely good at:

• Recalling intrinsic names and semantics (better than I am)
• Translating a working NEON kernel to AVX2, or to RVV
• Writing the scalar reference implementation and the test harness
• Explaining an unfamiliar instruction

What actually works: close the loop

Help your agent.

1. A benchmark it can run, over realistic inputs
2. A differential fuzzer against a scalar reference
3. llvm-mca (or perf) so it can see cycles, not vibes

Then let it iterate.

Case study: perfect hashing

Looking up!

• Map a string to a value: HTTP method, header name, URL scheme, keyword, MIME type
• The keys are fixed when you write the code
• Yet we hash, mask, probe, chase a pointer, compare: std::unordered_map, 12.8 ns
• With the keys known at compile time, we can build a perfect hash: no collisions, one candidate per slot

Library: github.com/ConstexprCore/perfect_hash (with Francisco Geiman Thiesen at Microsoft)

Case study: base64

Base64

• Encodes binary data as text using 64 characters (A-Z, a-z, 0-9, +, /)

• 3 bytes input → 4 characters output (33% overhead)

• Data URLs, email, JWTs, web APIs, embedded images

• "Hello, World!" → SGVsbG8sIFdvcmxkIQ==

Bit manipulation on a fixed schedule: the ideal SIMD problem.

New JavaScript functions

const b64 = Uint8Array.prototype.toBase64(bytes);
const recovered = Uint8Array.fromBase64(b64);

Test in your browser: https://simdutf.github.io/browserbase64/

The web's Base64 is not RFC 4648

• ES2026 Uint8Array.fromBase64 follows WHATWG forgiving base64: ignore ASCII whitespace, reject anything else
• MIME e-mail breaks lines every 76 characters; DNS zone files put spaces every 50; JSON and data URLs wrap freely
• Every earlier SIMD decoder assumed clean input: one \n and it falls back to a byte loop

Keep the SIMD throughput on input that has whitespace in it.

AVX-512 base64 encoding

• Encoding a 64-byte block requires only two non-memory instructions:
  vpermb (twice) and vpmultishiftqb.
• The right instruction turns an algorithm into a lookup.